feat(chat): add DB schema and Pydantic models for multi-model answers - 1/8

Add the foundational schema and type changes needed for multi-model
answer generation, where users can compare responses from up to 3 LLMs
side-by-side.

- Alembic migration: add `preferred_response_id` FK and
  `model_display_name` columns to `chat_message`
- Extend `SendMessageRequest` with `llm_overrides: list[LLMOverride]`
- Add `model_index` to `Placement` for streaming packet routing
- New `MultiModelMessageResponseIDInfo` packet type
- Extend `ChatMessageDetail` with `preferred_response_id` and
  `model_display_name`

.github/workflows/deployment.yml

@@ -29,15 +29,27 @@ jobs:
       build-backend-craft: ${{ steps.check.outputs.build-backend-craft }}
       build-model-server: ${{ steps.check.outputs.build-model-server }}
       is-cloud-tag: ${{ steps.check.outputs.is-cloud-tag }}
-      is-stable: ${{ steps.check.outputs.is-stable }}
       is-beta: ${{ steps.check.outputs.is-beta }}
-      is-stable-standalone: ${{ steps.check.outputs.is-stable-standalone }}
       is-beta-standalone: ${{ steps.check.outputs.is-beta-standalone }}
       is-craft-latest: ${{ steps.check.outputs.is-craft-latest }}
+      is-latest: ${{ steps.check.outputs.is-latest }}
       is-test-run: ${{ steps.check.outputs.is-test-run }}
       sanitized-tag: ${{ steps.check.outputs.sanitized-tag }}
       short-sha: ${{ steps.check.outputs.short-sha }}
     steps:
+      - name: Checkout (for git tags)
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # ratchet:actions/checkout@v6
+        with:
+          persist-credentials: false
+          fetch-depth: 0
+          fetch-tags: true
+
+      - name: Setup uv
+        uses: astral-sh/setup-uv@5a095e7a2014a4212f075830d4f7277575a9d098 # ratchet:astral-sh/setup-uv@v7
+        with:
+          version: "0.9.9"
+          enable-cache: false
+
       - name: Check which components to build and version info
         id: check
         env:
@@ -54,9 +66,9 @@ jobs:
           IS_VERSION_TAG=false
           IS_STABLE=false
           IS_BETA=false
-          IS_STABLE_STANDALONE=false
           IS_BETA_STANDALONE=false
           IS_CRAFT_LATEST=false
+          IS_LATEST=false
           IS_PROD_TAG=false
           IS_TEST_RUN=false
           BUILD_DESKTOP=false
@@ -104,13 +116,22 @@ jobs:
           fi
 
           # Standalone version checks (for backend/model-server - version excluding cloud tags)
-          if [[ "$IS_STABLE" == "true" ]] && [[ "$IS_CLOUD" != "true" ]]; then
-            IS_STABLE_STANDALONE=true
-          fi
           if [[ "$IS_BETA" == "true" ]] && [[ "$IS_CLOUD" != "true" ]]; then
             IS_BETA_STANDALONE=true
           fi
 
+          # Determine if this tag should get the "latest" Docker tag.
+          # Only the highest semver stable tag (vX.Y.Z exactly) gets "latest".
+          if [[ "$IS_STABLE" == "true" ]]; then
+            HIGHEST_STABLE=$(uv run --no-sync --with onyx-devtools ods latest-stable-tag) || {
+              echo "::error::Failed to determine highest stable tag via 'ods latest-stable-tag'"
+              exit 1
+            }
+            if [[ "$TAG" == "$HIGHEST_STABLE" ]]; then
+              IS_LATEST=true
+            fi
+          fi
+
           # Determine if this is a production tag
           # Production tags are: version tags (v1.2.3*) or nightly tags
           if [[ "$IS_VERSION_TAG" == "true" ]] || [[ "$IS_NIGHTLY" == "true" ]]; then
@@ -129,11 +150,10 @@ jobs:
             echo "build-backend-craft=$BUILD_BACKEND_CRAFT"
             echo "build-model-server=$BUILD_MODEL_SERVER"
             echo "is-cloud-tag=$IS_CLOUD"
-            echo "is-stable=$IS_STABLE"
             echo "is-beta=$IS_BETA"
-            echo "is-stable-standalone=$IS_STABLE_STANDALONE"
             echo "is-beta-standalone=$IS_BETA_STANDALONE"
             echo "is-craft-latest=$IS_CRAFT_LATEST"
+            echo "is-latest=$IS_LATEST"
             echo "is-test-run=$IS_TEST_RUN"
             echo "sanitized-tag=$SANITIZED_TAG"
             echo "short-sha=$SHORT_SHA"
@@ -600,7 +620,7 @@ jobs:
             latest=false
           tags: |
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run == 'true' && format('web-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
-            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-stable == 'true' && 'latest' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-latest == 'true' && 'latest' || '' }}
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && env.EDGE_TAG == 'true' && 'edge' || '' }}
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-beta == 'true' && 'beta' || '' }}
 
@@ -1037,7 +1057,7 @@ jobs:
             latest=false
           tags: |
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run == 'true' && format('backend-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
-            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-stable-standalone == 'true' && 'latest' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-latest == 'true' && 'latest' || '' }}
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && env.EDGE_TAG == 'true' && 'edge' || '' }}
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-beta-standalone == 'true' && 'beta' || '' }}
 
@@ -1473,7 +1493,7 @@ jobs:
             latest=false
           tags: |
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run == 'true' && format('model-server-{0}', needs.determine-builds.outputs.sanitized-tag) || github.ref_name }}
-            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-stable-standalone == 'true' && 'latest' || '' }}
+            type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-latest == 'true' && 'latest' || '' }}
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && env.EDGE_TAG == 'true' && 'edge' || '' }}
             type=raw,value=${{ needs.determine-builds.outputs.is-test-run != 'true' && needs.determine-builds.outputs.is-beta-standalone == 'true' && 'beta' || '' }}
 

backend/alembic/versions/a3f8b2c1d4e5_add_preferred_response_id_to_chat_message.py

@@ -0,0 +1,42 @@
+"""add multi-model columns to chat_message
+
+Revision ID: a3f8b2c1d4e5
+Revises: 27fb147a843f
+Create Date: 2026-03-12 10:00:00.000000
+
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = "a3f8b2c1d4e5"
+down_revision = "27fb147a843f"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "chat_message",
+        sa.Column(
+            "preferred_response_id",
+            sa.Integer(),
+            sa.ForeignKey("chat_message.id"),
+            nullable=True,
+        ),
+    )
+    op.add_column(
+        "chat_message",
+        sa.Column(
+            "model_display_name",
+            sa.String(),
+            nullable=True,
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("chat_message", "model_display_name")
+    op.drop_column("chat_message", "preferred_response_id")

backend/onyx/chat/models.py

@@ -8,6 +8,7 @@ from onyx.configs.constants import MessageType
 from onyx.context.search.models import SearchDoc
 from onyx.file_store.models import InMemoryChatFile
 from onyx.server.query_and_chat.models import MessageResponseIDInfo
+from onyx.server.query_and_chat.models import MultiModelMessageResponseIDInfo
 from onyx.server.query_and_chat.streaming_models import CitationInfo
 from onyx.server.query_and_chat.streaming_models import GeneratedImage
 from onyx.server.query_and_chat.streaming_models import Packet
@@ -35,7 +36,13 @@ class CreateChatSessionID(BaseModel):
     chat_session_id: UUID
 
 
-AnswerStreamPart = Packet | MessageResponseIDInfo | StreamingError | CreateChatSessionID
+AnswerStreamPart = (
+    Packet
+    | MessageResponseIDInfo
+    | MultiModelMessageResponseIDInfo
+    | StreamingError
+    | CreateChatSessionID
+)
 
 AnswerStream = Iterator[AnswerStreamPart]
 

backend/onyx/db/models.py

@@ -2622,6 +2622,18 @@ class ChatMessage(Base):
         ForeignKey("chat_message.id"), nullable=True
     )
 
+    # For multi-model turns: the user message points to which assistant response
+    # was selected as the preferred one to continue the conversation with.
+    # Only set on user messages that triggered a multi-model generation.
+    preferred_response_id: Mapped[int | None] = mapped_column(
+        ForeignKey("chat_message.id"), nullable=True
+    )
+
+    # The display name of the model that generated this assistant message
+    # (e.g. "GPT-4", "Claude Opus"). Used on session reload to label
+    # multi-model response panels and <> navigation arrows.
+    model_display_name: Mapped[str | None] = mapped_column(String, nullable=True)
+
     # Only set on summary messages - the ID of the last message included in this summary
     # Used for chat history compression
     last_summarized_message_id: Mapped[int | None] = mapped_column(
@@ -2696,6 +2708,12 @@ class ChatMessage(Base):
         remote_side="ChatMessage.id",
     )
 
+    preferred_response: Mapped["ChatMessage | None"] = relationship(
+        "ChatMessage",
+        foreign_keys=[preferred_response_id],
+        remote_side="ChatMessage.id",
+    )
+
     # Chat messages only need to know their immediate tool call children
     # If there are nested tool calls, they are stored in the tool_call_children relationship.
     tool_calls: Mapped[list["ToolCall"] | None] = relationship(

backend/onyx/onyxbot/slack/blocks.py

@@ -1,6 +1,3 @@
-import logging
-import re
-from dataclasses import dataclass
 from datetime import datetime
 from typing import cast
 
@@ -49,8 +46,6 @@ from onyx.onyxbot.slack.utils import remove_slack_text_interactions
 from onyx.onyxbot.slack.utils import translate_vespa_highlight_to_slack
 from onyx.utils.text_processing import decode_escapes
 
-logger = logging.getLogger(__name__)
-
 _MAX_BLURB_LEN = 45
 
 
@@ -81,99 +76,6 @@ def get_feedback_reminder_blocks(thread_link: str, include_followup: bool) -> Bl
     return SectionBlock(text=text)
 
 
-@dataclass
-class CodeSnippet:
-    """A code block extracted from the answer to be uploaded as a Slack file."""
-
-    code: str
-    language: str
-    filename: str
-
-
-_SECTION_BLOCK_LIMIT = 3000
-
-# Matches fenced code blocks: ```lang\n...\n```
-# The opening fence must start at the beginning of a line (^), may have an
-# optional language specifier (\w*), followed by a newline. The closing fence
-# is ``` at the beginning of a line. We also handle an optional trailing
-# newline on the closing fence line to be robust against different formatting.
-_CODE_FENCE_RE = re.compile(
-    r"^```(\w*)\n(.*?)^```\s*$",
-    re.MULTILINE | re.DOTALL,
-)
-
-
-def _extract_code_snippets(
-    text: str, limit: int = _SECTION_BLOCK_LIMIT
-) -> tuple[str, list[CodeSnippet]]:
-    """Extract code blocks that would push the text over *limit*.
-
-    Returns (cleaned_text, snippets) where *cleaned_text* has large code
-    blocks replaced with a placeholder and *snippets* contains the extracted
-    code to be uploaded as Slack file snippets.
-
-    Uses a two-pass approach: first collect all matches, then decide which
-    to extract based on cumulative removal so each decision accounts for
-    previously extracted blocks.
-
-    Pass *limit=0* to force-extract ALL code blocks unconditionally.
-    """
-    if limit > 0 and len(text) <= limit:
-        return text, []
-
-    # Pass 1: collect all code-fence matches
-    matches = list(_CODE_FENCE_RE.finditer(text))
-    if not matches:
-        return text, []
-
-    # Pass 2: decide which blocks to extract, accounting for cumulative removal.
-    # Only extract if the text is still over the limit OR the block is very large.
-    # With limit=0, extract everything unconditionally.
-    extract_indices: set[int] = set()
-    removed_chars = 0
-    for i, match in enumerate(matches):
-        full_block = match.group(0)
-        if limit == 0:
-            extract_indices.add(i)
-            removed_chars += len(full_block)
-        else:
-            current_len = len(text) - removed_chars
-            if current_len > limit and current_len - len(full_block) <= limit:
-                extract_indices.add(i)
-                removed_chars += len(full_block)
-            elif len(full_block) > limit // 2:
-                extract_indices.add(i)
-                removed_chars += len(full_block)
-
-    if not extract_indices:
-        return text, []
-
-    # Build cleaned text and snippets by processing matches in reverse
-    # so character offsets remain valid.
-    snippets: list[CodeSnippet] = []
-    cleaned = text
-    for i in sorted(extract_indices, reverse=True):
-        match = matches[i]
-        lang = match.group(1) or ""
-        code = match.group(2)
-        ext = lang if lang else "txt"
-        snippets.append(
-            CodeSnippet(
-                code=code.strip(),
-                language=lang or "text",
-                filename=f"code_{len(extract_indices) - len(snippets)}.{ext}",
-            )
-        )
-        cleaned = cleaned[: match.start()] + cleaned[match.end() :]
-
-    # Snippets were appended in reverse order — flip to match document order
-    snippets.reverse()
-
-    # Clean up any triple+ blank lines left by extraction
-    cleaned = re.sub(r"\n{3,}", "\n\n", cleaned).strip()
-    return cleaned, snippets
-
-
 def _split_text(text: str, limit: int = 3000) -> list[str]:
     if len(text) <= limit:
         return [text]
@@ -515,7 +417,7 @@ def _build_citations_blocks(
 
 def _build_main_response_blocks(
     answer: ChatBasicResponse,
-) -> tuple[list[Block], list[CodeSnippet]]:
+) -> list[Block]:
     # TODO: add back in later when auto-filtering is implemented
     # if (
     #     retrieval_info.applied_time_cutoff
@@ -546,45 +448,9 @@ def _build_main_response_blocks(
     # replaces markdown links with slack format links
     formatted_answer = format_slack_message(answer.answer)
     answer_processed = decode_escapes(remove_slack_text_interactions(formatted_answer))
+    answer_blocks = [SectionBlock(text=text) for text in _split_text(answer_processed)]
 
-    # Extract large code blocks as snippets to upload separately,
-    # avoiding broken code fences when splitting across SectionBlocks.
-    cleaned_text, code_snippets = _extract_code_snippets(answer_processed)
-    logger.info(
-        "Code extraction: input=%d chars, cleaned=%d chars, snippets=%d",
-        len(answer_processed),
-        len(cleaned_text),
-        len(code_snippets),
-    )
-
-    if len(cleaned_text) <= _SECTION_BLOCK_LIMIT:
-        answer_blocks = [SectionBlock(text=cleaned_text)]
-    elif "```" not in cleaned_text:
-        # No code fences — safe to split at word boundaries.
-        answer_blocks = [
-            SectionBlock(text=text)
-            for text in _split_text(cleaned_text, limit=_SECTION_BLOCK_LIMIT)
-        ]
-    else:
-        # Text still has code fences after extraction and exceeds the
-        # SectionBlock limit. Splitting would break the fences, so fall
-        # back to uploading the entire remaining code as another snippet
-        # and keeping only the prose in the blocks.
-        logger.warning(
-            "Cleaned text still has code fences (%d chars); "
-            "force-extracting remaining code blocks",
-            len(cleaned_text),
-        )
-        remaining_cleaned, remaining_snippets = _extract_code_snippets(
-            cleaned_text, limit=0
-        )
-        code_snippets.extend(remaining_snippets)
-        answer_blocks = [
-            SectionBlock(text=text)
-            for text in _split_text(remaining_cleaned, limit=_SECTION_BLOCK_LIMIT)
-        ]
-
-    return cast(list[Block], answer_blocks), code_snippets
+    return cast(list[Block], answer_blocks)
 
 
 def _build_continue_in_web_ui_block(
@@ -665,13 +531,10 @@ def build_slack_response_blocks(
     skip_ai_feedback: bool = False,
     offer_ephemeral_publication: bool = False,
     skip_restated_question: bool = False,
-) -> tuple[list[Block], list[CodeSnippet]]:
+) -> list[Block]:
     """
     This function is a top level function that builds all the blocks for the Slack response.
     It also handles combining all the blocks together.
-
-    Returns (blocks, code_snippets) where code_snippets should be uploaded
-    as Slack file snippets in the same thread.
     """
     # If called with the OnyxBot slash command, the question is lost so we have to reshow it
     if not skip_restated_question:
@@ -681,7 +544,7 @@ def build_slack_response_blocks(
     else:
         restate_question_block = []
 
-    answer_blocks, code_snippets = _build_main_response_blocks(answer)
+    answer_blocks = _build_main_response_blocks(answer)
 
     web_follow_up_block = []
     if channel_conf and channel_conf.get("show_continue_in_web_ui"):
@@ -747,4 +610,4 @@ def build_slack_response_blocks(
         + follow_up_block
     )
 
-    return all_blocks, code_snippets
+    return all_blocks

backend/onyx/onyxbot/slack/handlers/handle_buttons.py

@@ -282,7 +282,7 @@ def handle_publish_ephemeral_message_button(
             logger.error(f"Failed to send webhook: {e}")
 
         # remove handling of empheremal block and add AI feedback.
-        all_blocks, _ = build_slack_response_blocks(
+        all_blocks = build_slack_response_blocks(
             answer=onyx_bot_answer,
             message_info=slack_message_info,
             channel_conf=channel_conf,
@@ -311,7 +311,7 @@ def handle_publish_ephemeral_message_button(
     elif action_id == KEEP_TO_YOURSELF_ACTION_ID:
         # Keep as ephemeral message in channel or thread, but remove the publish button and add feedback button
 
-        changed_blocks, _ = build_slack_response_blocks(
+        changed_blocks = build_slack_response_blocks(
             answer=onyx_bot_answer,
             message_info=slack_message_info,
             channel_conf=channel_conf,

backend/onyx/onyxbot/slack/handlers/handle_regular_answer.py

@@ -25,7 +25,6 @@ from onyx.db.models import User
 from onyx.db.persona import get_persona_by_id
 from onyx.db.users import get_user_by_email
 from onyx.onyxbot.slack.blocks import build_slack_response_blocks
-from onyx.onyxbot.slack.blocks import CodeSnippet
 from onyx.onyxbot.slack.constants import SLACK_CHANNEL_REF_PATTERN
 from onyx.onyxbot.slack.handlers.utils import send_team_member_message
 from onyx.onyxbot.slack.models import SlackMessageInfo
@@ -135,65 +134,6 @@ def build_slack_context_str(
     return slack_context_str + "\n\n".join(message_strs)
 
 
-# Normalize common LLM language aliases to Slack's expected snippet_type values.
-# Slack silently falls back to plain text for unrecognized types, so this map
-# only needs to cover the most common mismatches.
-_SNIPPET_TYPE_MAP: dict[str, str] = {
-    "py": "python",
-    "js": "javascript",
-    "ts": "typescript",
-    "tsx": "typescript",
-    "jsx": "javascript",
-    "sh": "shell",
-    "bash": "shell",
-    "zsh": "shell",
-    "yml": "yaml",
-    "rb": "ruby",
-    "rs": "rust",
-    "cs": "csharp",
-    "md": "markdown",
-    "txt": "text",
-    "text": "plain_text",
-}
-
-
-def _upload_code_snippets(
-    client: WebClient,
-    channel: str,
-    thread_ts: str,
-    snippets: list[CodeSnippet],
-    logger: OnyxLoggingAdapter,
-    receiver_ids: list[str] | None = None,
-    send_as_ephemeral: bool | None = None,
-) -> None:
-    """Upload extracted code blocks as Slack file snippets in the thread."""
-    for snippet in snippets:
-        try:
-            snippet_type = _SNIPPET_TYPE_MAP.get(snippet.language, snippet.language)
-            client.files_upload_v2(
-                channel=channel,
-                thread_ts=thread_ts,
-                content=snippet.code,
-                filename=snippet.filename,
-                snippet_type=snippet_type,
-            )
-        except Exception:
-            logger.warning(
-                f"Failed to upload code snippet {snippet.filename}, "
-                "falling back to inline code block"
-            )
-            # Fall back to posting as a regular message with code fences,
-            # preserving the same visibility as the primary response.
-            respond_in_thread_or_channel(
-                client=client,
-                channel=channel,
-                receiver_ids=receiver_ids,
-                text=f"```{snippet.language}\n{snippet.code}\n```",
-                thread_ts=thread_ts,
-                send_as_ephemeral=send_as_ephemeral,
-            )
-
-
 def handle_regular_answer(
     message_info: SlackMessageInfo,
     slack_channel_config: SlackChannelConfig,
@@ -447,7 +387,7 @@ def handle_regular_answer(
         offer_ephemeral_publication = False
         skip_ai_feedback = False
 
-    all_blocks, code_snippets = build_slack_response_blocks(
+    all_blocks = build_slack_response_blocks(
         message_info=message_info,
         answer=answer,
         channel_conf=channel_conf,
@@ -474,20 +414,6 @@ def handle_regular_answer(
             send_as_ephemeral=send_as_ephemeral,
         )
 
-        # Upload extracted code blocks as Slack file snippets so they
-        # render as collapsible, syntax-highlighted blocks in the thread.
-        snippet_thread_ts = target_thread_ts or message_ts_to_respond_to
-        if code_snippets and snippet_thread_ts:
-            _upload_code_snippets(
-                client=client,
-                channel=channel,
-                thread_ts=snippet_thread_ts,
-                snippets=code_snippets,
-                logger=logger,
-                receiver_ids=target_receiver_ids,
-                send_as_ephemeral=send_as_ephemeral,
-            )
-
         # For DM (ephemeral message), we need to create a thread via a normal message so the user can see
         # the ephemeral message. This also will give the user a notification which ephemeral message does not.
         # if there is no message_ts_to_respond_to, and we have made it this far, then this is a /onyx message

backend/onyx/server/query_and_chat/models.py

@@ -41,6 +41,16 @@ class MessageResponseIDInfo(BaseModel):
     reserved_assistant_message_id: int
 
 
+class MultiModelMessageResponseIDInfo(BaseModel):
+    """Sent at the start of a multi-model streaming response.
+    Contains the user message ID and the reserved assistant message IDs
+    for each model being run in parallel."""
+
+    user_message_id: int | None
+    reserved_assistant_message_ids: list[int]
+    model_names: list[str]
+
+
 class SourceTag(Tag):
     source: DocumentSource
 
@@ -86,6 +96,10 @@ class SendMessageRequest(BaseModel):
     message: str
 
     llm_override: LLMOverride | None = None
+    # For multi-model mode: up to 3 LLM overrides to run in parallel.
+    # When provided with >1 entry, triggers multi-model streaming.
+    # Backward-compat: if only `llm_override` is set, single-model path is used.
+    llm_overrides: list[LLMOverride] | None = None
     # Test-only override for deterministic LiteLLM mock responses.
     mock_llm_response: str | None = None
 
@@ -211,6 +225,10 @@ class ChatMessageDetail(BaseModel):
     error: str | None = None
     current_feedback: str | None = None  # "like" | "dislike" | null
     processing_duration_seconds: float | None = None
+    # For multi-model turns: the preferred assistant response ID (set on user messages only)
+    preferred_response_id: int | None = None
+    # The display name of the model that generated this message (e.g. "GPT-4", "Claude Opus")
+    model_display_name: str | None = None
 
     def model_dump(self, *args: list, **kwargs: dict[str, Any]) -> dict[str, Any]:  # type: ignore
         initial_dict = super().model_dump(mode="json", *args, **kwargs)  # type: ignore

backend/onyx/server/query_and_chat/placement.py

@@ -8,3 +8,6 @@ class Placement(BaseModel):
     tab_index: int = 0
     # Used for tools/agents that call other tools, this currently doesn't support nested agents but can be added later
     sub_turn_index: int | None = None
+    # For multi-model streaming: identifies which model (0, 1, 2) this packet belongs to.
+    # None for single-model (default) responses.
+    model_index: int | None = None

backend/tests/unit/onyx/chat/test_multi_model_types.py

@@ -0,0 +1,112 @@
+"""Unit tests for multi-model schema and Pydantic model additions."""
+
+from datetime import datetime
+
+from onyx.configs.constants import MessageType
+from onyx.llm.override_models import LLMOverride
+from onyx.server.query_and_chat.models import ChatMessageDetail
+from onyx.server.query_and_chat.models import MultiModelMessageResponseIDInfo
+from onyx.server.query_and_chat.models import SendMessageRequest
+from onyx.server.query_and_chat.placement import Placement
+
+
+def test_placement_model_index_default_none() -> None:
+    p = Placement(turn_index=0)
+    assert p.model_index is None
+
+
+def test_placement_model_index_set() -> None:
+    p = Placement(turn_index=0, model_index=2)
+    assert p.model_index == 2
+
+
+def test_placement_serialization_with_model_index() -> None:
+    p = Placement(turn_index=1, tab_index=0, model_index=1)
+    data = p.model_dump()
+    assert data["model_index"] == 1
+    restored = Placement(**data)
+    assert restored.model_index == 1
+
+
+def test_multi_model_message_response_id_info() -> None:
+    info = MultiModelMessageResponseIDInfo(
+        user_message_id=42,
+        reserved_assistant_message_ids=[100, 101, 102],
+        model_names=["gpt-4", "claude-3-opus", "gemini-pro"],
+    )
+    data = info.model_dump()
+    assert data["user_message_id"] == 42
+    assert len(data["reserved_assistant_message_ids"]) == 3
+    assert len(data["model_names"]) == 3
+
+
+def test_multi_model_message_response_id_info_null_user() -> None:
+    info = MultiModelMessageResponseIDInfo(
+        user_message_id=None,
+        reserved_assistant_message_ids=[10],
+        model_names=["gpt-4"],
+    )
+    assert info.user_message_id is None
+
+
+def test_send_message_request_llm_overrides_none_by_default() -> None:
+    req = SendMessageRequest(
+        message="hello",
+        chat_session_id="00000000-0000-0000-0000-000000000001",
+    )
+    assert req.llm_overrides is None
+    assert req.llm_override is None
+
+
+def test_send_message_request_with_llm_overrides() -> None:
+    overrides = [
+        LLMOverride(model_provider="openai", model_version="gpt-4"),
+        LLMOverride(model_provider="anthropic", model_version="claude-3-opus"),
+    ]
+    req = SendMessageRequest(
+        message="compare these",
+        chat_session_id="00000000-0000-0000-0000-000000000001",
+        llm_overrides=overrides,
+    )
+    assert req.llm_overrides is not None
+    assert len(req.llm_overrides) == 2
+
+
+def test_send_message_request_backward_compat_single_override() -> None:
+    """Existing single llm_override still works alongside new llm_overrides field."""
+    req = SendMessageRequest(
+        message="single model",
+        chat_session_id="00000000-0000-0000-0000-000000000001",
+        llm_override=LLMOverride(model_provider="openai", model_version="gpt-4"),
+    )
+    assert req.llm_override is not None
+    assert req.llm_overrides is None
+
+
+def test_chat_message_detail_multi_model_fields_default_none() -> None:
+    detail = ChatMessageDetail(
+        message_id=1,
+        message="hello",
+        message_type=MessageType.USER,
+        time_sent=datetime.now(),
+        files=[],
+    )
+    assert detail.preferred_response_id is None
+    assert detail.model_display_name is None
+
+
+def test_chat_message_detail_multi_model_fields_set() -> None:
+    detail = ChatMessageDetail(
+        message_id=1,
+        message="response from gpt-4",
+        message_type=MessageType.ASSISTANT,
+        time_sent=datetime.now(),
+        files=[],
+        preferred_response_id=42,
+        model_display_name="GPT-4",
+    )
+    assert detail.preferred_response_id == 42
+    assert detail.model_display_name == "GPT-4"
+    data = detail.model_dump()
+    assert data["preferred_response_id"] == 42
+    assert data["model_display_name"] == "GPT-4"

backend/tests/unit/onyx/onyxbot/test_slack_blocks.py

@@ -7,9 +7,6 @@ import timeago  # type: ignore
 from onyx.configs.constants import DocumentSource
 from onyx.context.search.models import SavedSearchDoc
 from onyx.onyxbot.slack.blocks import _build_documents_blocks
-from onyx.onyxbot.slack.blocks import _extract_code_snippets
-from onyx.onyxbot.slack.blocks import _split_text
-from onyx.onyxbot.slack.handlers.handle_regular_answer import _SNIPPET_TYPE_MAP
 
 
 def _make_saved_doc(updated_at: datetime | None) -> SavedSearchDoc:
@@ -72,148 +69,3 @@ def test_build_documents_blocks_formats_naive_timestamp(
     formatted_timestamp: datetime = captured["doc"]
     expected_timestamp: datetime = naive_timestamp.replace(tzinfo=pytz.utc)
     assert formatted_timestamp == expected_timestamp
-
-
-# ---------------------------------------------------------------------------
-# _split_text tests
-# ---------------------------------------------------------------------------
-
-
-class TestSplitText:
-    def test_short_text_returns_single_chunk(self) -> None:
-        result = _split_text("hello world", limit=100)
-        assert result == ["hello world"]
-
-    def test_splits_at_space_boundary(self) -> None:
-        text = "aaa bbb ccc ddd"
-        result = _split_text(text, limit=8)
-        assert len(result) >= 2
-
-    def test_no_code_fences_splits_normally(self) -> None:
-        text = "word " * 100  # 500 chars
-        result = _split_text(text, limit=100)
-        assert len(result) >= 5
-        for chunk in result:
-            assert "```" not in chunk
-
-
-# ---------------------------------------------------------------------------
-# _extract_code_snippets tests
-# ---------------------------------------------------------------------------
-
-
-class TestExtractCodeSnippets:
-    def test_short_text_no_extraction(self) -> None:
-        text = "short answer with ```python\nprint('hi')\n``` inline"
-        cleaned, snippets = _extract_code_snippets(text, limit=3000)
-        assert cleaned == text
-        assert snippets == []
-
-    def test_large_code_block_extracted(self) -> None:
-        code = "x = 1\n" * 200  # ~1200 chars of code
-        text = f"Here is the solution:\n```python\n{code}```\nHope that helps!"
-        cleaned, snippets = _extract_code_snippets(text, limit=200)
-
-        assert len(snippets) == 1
-        assert snippets[0].language == "python"
-        assert snippets[0].filename == "code_1.python"
-        assert "x = 1" in snippets[0].code
-        # Code block should be removed from cleaned text
-        assert "```" not in cleaned
-        assert "Here is the solution" in cleaned
-        assert "Hope that helps!" in cleaned
-
-    def test_multiple_code_blocks_only_large_ones_extracted(self) -> None:
-        small_code = "print('hi')"
-        large_code = "x = 1\n" * 300
-        text = (
-            f"First:\n```python\n{small_code}\n```\n"
-            f"Second:\n```javascript\n{large_code}\n```\n"
-            "Done!"
-        )
-        cleaned, snippets = _extract_code_snippets(text, limit=500)
-
-        # The large block should be extracted
-        assert len(snippets) >= 1
-        langs = [s.language for s in snippets]
-        assert "javascript" in langs
-
-    def test_language_specifier_captured(self) -> None:
-        code = "fn main() {}\n" * 100
-        text = f"```rust\n{code}```"
-        _, snippets = _extract_code_snippets(text, limit=100)
-
-        assert len(snippets) == 1
-        assert snippets[0].language == "rust"
-        assert snippets[0].filename == "code_1.rust"
-
-    def test_no_language_defaults_to_text(self) -> None:
-        code = "some output\n" * 100
-        text = f"```\n{code}```"
-        _, snippets = _extract_code_snippets(text, limit=100)
-
-        assert len(snippets) == 1
-        assert snippets[0].language == "text"
-        assert snippets[0].filename == "code_1.txt"
-
-    def test_cleaned_text_has_no_triple_blank_lines(self) -> None:
-        code = "x = 1\n" * 200
-        text = f"Before\n\n```python\n{code}```\n\nAfter"
-        cleaned, _ = _extract_code_snippets(text, limit=100)
-
-        assert "\n\n\n" not in cleaned
-
-    def test_multiple_blocks_cumulative_removal(self) -> None:
-        """When multiple code blocks exist, extraction decisions should
-        account for previously extracted blocks (two-pass logic).
-        Blocks must be smaller than limit//2 so the 'very large block'
-        override doesn't trigger — we're testing the cumulative logic only."""
-        # Each fenced block is ~103 chars (```python\n + 15*6 + ```)
-        block_a = "a = 1\n" * 15  # 90 chars of code
-        block_b = "b = 2\n" * 15  # 90 chars of code
-        prose = "x" * 200
-        # Total: ~200 + 103 + 103 + overhead ≈ 420 chars
-        # limit=300, limit//2=150. Each block (~103) < 150, so only
-        # the cumulative check applies. Removing block_a (~103 chars)
-        # brings us to ~317 > 300, so block_b should also be extracted.
-        # But with limit=350: removing block_a → ~317 ≤ 350, stop.
-        text = f"{prose}\n```python\n{block_a}```\n```python\n{block_b}```\nEnd"
-        cleaned, snippets = _extract_code_snippets(text, limit=350)
-
-        # After extracting block_a the text is ≤ 350, so block_b stays.
-        assert len(snippets) == 1
-        assert snippets[0].filename == "code_1.python"
-        # block_b should still be in the cleaned text
-        assert "b = 2" in cleaned
-
-
-# ---------------------------------------------------------------------------
-# _SNIPPET_TYPE_MAP tests
-# ---------------------------------------------------------------------------
-
-
-class TestSnippetTypeMap:
-    @pytest.mark.parametrize(
-        "alias,expected",
-        [
-            ("py", "python"),
-            ("js", "javascript"),
-            ("ts", "typescript"),
-            ("tsx", "typescript"),
-            ("jsx", "javascript"),
-            ("sh", "shell"),
-            ("bash", "shell"),
-            ("yml", "yaml"),
-            ("rb", "ruby"),
-            ("rs", "rust"),
-            ("cs", "csharp"),
-            ("md", "markdown"),
-            ("text", "plain_text"),
-        ],
-    )
-    def test_common_aliases_normalized(self, alias: str, expected: str) -> None:
-        assert _SNIPPET_TYPE_MAP[alias] == expected
-
-    def test_unknown_language_passes_through(self) -> None:
-        unknown = "haskell"
-        assert _SNIPPET_TYPE_MAP.get(unknown, unknown) == "haskell"

web/src/hooks/useAdminUsers.ts

@@ -1,5 +1,6 @@
 "use client";
 
+import { useCallback } from "react";
 import useSWR from "swr";
 import { errorHandlingFetcher } from "@/lib/fetcher";
 import { NEXT_PUBLIC_CLOUD_ENABLED } from "@/lib/constants";
@@ -112,11 +113,11 @@ export default function useAdminUsers() {
   const isLoading = acceptedLoading || invitedLoading || requestedLoading;
   const error = acceptedError ?? invitedError ?? requestedError;
 
-  function refresh() {
+  const refresh = useCallback(() => {
     acceptedMutate();
     invitedMutate();
     requestedMutate();
-  }
+  }, [acceptedMutate, invitedMutate, requestedMutate]);
 
   return { users, isLoading, error, refresh };
 }

web/src/refresh-components/inputs/InputSelect.tsx

@@ -2,7 +2,7 @@
 
 import * as React from "react";
 import * as SelectPrimitive from "@radix-ui/react-select";
-import { cn, noProp } from "@/lib/utils";
+import { cn } from "@/lib/utils";
 import LineItem, { LineItemProps } from "@/refresh-components/buttons/LineItem";
 import Text from "@/refresh-components/texts/Text";
 import type { IconProps } from "@opal/types";
@@ -298,7 +298,10 @@ function InputSelectContent({
         )}
         sideOffset={4}
         position="popper"
-        onMouseDown={noProp()}
+        onMouseDown={(e) => {
+          e.stopPropagation();
+          e.preventDefault();
+        }}
         {...props}
       >
         <SelectPrimitive.Viewport className="flex flex-col gap-1">

web/src/refresh-pages/admin/UsersPage/UserRowActions.tsx

@@ -0,0 +1,210 @@
+"use client";
+
+import { useState } from "react";
+import { Button } from "@opal/components";
+import { SvgMoreHorizontal, SvgXCircle, SvgTrash, SvgCheck } from "@opal/icons";
+import { Disabled } from "@opal/core";
+import Popover from "@/refresh-components/Popover";
+import ConfirmationModalLayout from "@/refresh-components/layouts/ConfirmationModalLayout";
+import Text from "@/refresh-components/texts/Text";
+import { UserStatus } from "@/lib/types";
+import { toast } from "@/hooks/useToast";
+import { deactivateUser, activateUser, deleteUser } from "./svc";
+import type { UserRow } from "./interfaces";
+
+// ---------------------------------------------------------------------------
+// Types
+// ---------------------------------------------------------------------------
+
+type ModalType = "deactivate" | "activate" | "delete" | null;
+
+interface UserRowActionsProps {
+  user: UserRow;
+  onMutate: () => void;
+}
+
+// ---------------------------------------------------------------------------
+// Component
+// ---------------------------------------------------------------------------
+
+export default function UserRowActions({
+  user,
+  onMutate,
+}: UserRowActionsProps) {
+  const [modal, setModal] = useState<ModalType>(null);
+  const [popoverOpen, setPopoverOpen] = useState(false);
+  const [isSubmitting, setIsSubmitting] = useState(false);
+
+  async function handleAction(
+    action: () => Promise<void>,
+    successMessage: string
+  ) {
+    setIsSubmitting(true);
+    try {
+      await action();
+      onMutate();
+      toast.success(successMessage);
+      setModal(null);
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : "An error occurred");
+    } finally {
+      setIsSubmitting(false);
+    }
+  }
+
+  // Only show actions for accepted users (active or inactive).
+  // Invited/requested users have no row actions in this PR.
+  if (
+    user.status !== UserStatus.ACTIVE &&
+    user.status !== UserStatus.INACTIVE
+  ) {
+    return null;
+  }
+
+  // SCIM-managed users cannot be modified from the UI — changes would be
+  // overwritten on the next IdP sync.
+  if (user.is_scim_synced) {
+    return null;
+  }
+
+  return (
+    <>
+      <Popover open={popoverOpen} onOpenChange={setPopoverOpen}>
+        <Popover.Trigger asChild>
+          <Button prominence="tertiary" icon={SvgMoreHorizontal} />
+        </Popover.Trigger>
+        <Popover.Content align="end">
+          <div className="flex flex-col gap-0.5 p-1">
+            {user.status === UserStatus.ACTIVE ? (
+              <Button
+                prominence="tertiary"
+                icon={SvgXCircle}
+                onClick={() => {
+                  setPopoverOpen(false);
+                  setModal("deactivate");
+                }}
+              >
+                Deactivate User
+              </Button>
+            ) : (
+              <>
+                <Button
+                  prominence="tertiary"
+                  icon={SvgCheck}
+                  onClick={() => {
+                    setPopoverOpen(false);
+                    setModal("activate");
+                  }}
+                >
+                  Activate User
+                </Button>
+                <Button
+                  prominence="tertiary"
+                  variant="danger"
+                  icon={SvgTrash}
+                  onClick={() => {
+                    setPopoverOpen(false);
+                    setModal("delete");
+                  }}
+                >
+                  Delete User
+                </Button>
+              </>
+            )}
+          </div>
+        </Popover.Content>
+      </Popover>
+
+      {modal === "deactivate" && (
+        <ConfirmationModalLayout
+          icon={SvgXCircle}
+          title="Deactivate User"
+          onClose={isSubmitting ? undefined : () => setModal(null)}
+          submit={
+            <Disabled disabled={isSubmitting}>
+              <Button
+                variant="danger"
+                onClick={async () => {
+                  await handleAction(
+                    () => deactivateUser(user.email),
+                    "User deactivated"
+                  );
+                }}
+              >
+                Deactivate
+              </Button>
+            </Disabled>
+          }
+        >
+          <Text as="p" text03>
+            <Text as="span" text05>
+              {user.email}
+            </Text>{" "}
+            will immediately lose access to Onyx. Their sessions and agents will
+            be preserved. You can reactivate this account later.
+          </Text>
+        </ConfirmationModalLayout>
+      )}
+
+      {modal === "activate" && (
+        <ConfirmationModalLayout
+          icon={SvgCheck}
+          title="Activate User"
+          onClose={isSubmitting ? undefined : () => setModal(null)}
+          submit={
+            <Disabled disabled={isSubmitting}>
+              <Button
+                onClick={async () => {
+                  await handleAction(
+                    () => activateUser(user.email),
+                    "User activated"
+                  );
+                }}
+              >
+                Activate
+              </Button>
+            </Disabled>
+          }
+        >
+          <Text as="p" text03>
+            <Text as="span" text05>
+              {user.email}
+            </Text>{" "}
+            will regain access to Onyx.
+          </Text>
+        </ConfirmationModalLayout>
+      )}
+
+      {modal === "delete" && (
+        <ConfirmationModalLayout
+          icon={SvgTrash}
+          title="Delete User"
+          onClose={isSubmitting ? undefined : () => setModal(null)}
+          submit={
+            <Disabled disabled={isSubmitting}>
+              <Button
+                variant="danger"
+                onClick={async () => {
+                  await handleAction(
+                    () => deleteUser(user.email),
+                    "User deleted"
+                  );
+                }}
+              >
+                Delete
+              </Button>
+            </Disabled>
+          }
+        >
+          <Text as="p" text03>
+            <Text as="span" text05>
+              {user.email}
+            </Text>{" "}
+            will be permanently removed from Onyx. All of their session history
+            will be deleted. This cannot be undone.
+          </Text>
+        </ConfirmationModalLayout>
+      )}
+    </>
+  );
+}

web/src/refresh-pages/admin/UsersPage/UsersTable.tsx

@@ -21,6 +21,7 @@ import InputTypeIn from "@/refresh-components/inputs/InputTypeIn";
 import useAdminUsers from "@/hooks/useAdminUsers";
 import useGroups from "@/hooks/useGroups";
 import UserFilters from "./UserFilters";
+import UserRowActions from "./UserRowActions";
 import type {
   UserRow,
   UserGroupInfo,
@@ -133,50 +134,54 @@ function renderLastUpdatedColumn(value: string | null) {
 }
 
 // ---------------------------------------------------------------------------
-// Columns (stable reference — defined at module scope)
+// Columns
 // ---------------------------------------------------------------------------
 
 const tc = createTableColumns<UserRow>();
 
-const columns = [
-  tc.qualifier({
-    content: "avatar-user",
-    getInitials: (row) => getInitials(row.personal_name, row.email),
-    selectable: false,
-  }),
-  tc.column("email", {
-    header: "Name",
-    weight: 22,
-    minWidth: 140,
-    cell: renderNameColumn,
-  }),
-  tc.column("groups", {
-    header: "Groups",
-    weight: 24,
-    minWidth: 200,
-    enableSorting: false,
-    cell: renderGroupsColumn,
-  }),
-  tc.column("role", {
-    header: "Account Type",
-    weight: 16,
-    minWidth: 180,
-    cell: renderRoleColumn,
-  }),
-  tc.column("status", {
-    header: "Status",
-    weight: 14,
-    minWidth: 100,
-    cell: renderStatusColumn,
-  }),
-  tc.column("updated_at", {
-    header: "Last Updated",
-    weight: 14,
-    minWidth: 100,
-    cell: renderLastUpdatedColumn,
-  }),
-  tc.actions(),
-];
+function buildColumns(onMutate: () => void) {
+  return [
+    tc.qualifier({
+      content: "avatar-user",
+      getInitials: (row) => getInitials(row.personal_name, row.email),
+      selectable: false,
+    }),
+    tc.column("email", {
+      header: "Name",
+      weight: 22,
+      minWidth: 140,
+      cell: renderNameColumn,
+    }),
+    tc.column("groups", {
+      header: "Groups",
+      weight: 24,
+      minWidth: 200,
+      enableSorting: false,
+      cell: renderGroupsColumn,
+    }),
+    tc.column("role", {
+      header: "Account Type",
+      weight: 16,
+      minWidth: 180,
+      cell: renderRoleColumn,
+    }),
+    tc.column("status", {
+      header: "Status",
+      weight: 14,
+      minWidth: 100,
+      cell: renderStatusColumn,
+    }),
+    tc.column("updated_at", {
+      header: "Last Updated",
+      weight: 14,
+      minWidth: 100,
+      cell: renderLastUpdatedColumn,
+    }),
+    tc.actions({
+      cell: (row) => <UserRowActions user={row} onMutate={onMutate} />,
+    }),
+  ];
+}
 
 // ---------------------------------------------------------------------------
 // Component
@@ -213,7 +218,9 @@ export default function UsersTable({
     [allGroups]
   );
 
-  const { users, isLoading, error } = useAdminUsers();
+  const { users, isLoading, error, refresh } = useAdminUsers();
+
+  const columns = useMemo(() => buildColumns(refresh), [refresh]);
 
   // Client-side filtering
   const filteredUsers = useMemo(() => {

web/src/refresh-pages/admin/UsersPage/svc.ts

@@ -0,0 +1,44 @@
+async function parseErrorDetail(
+  res: Response,
+  fallback: string
+): Promise<string> {
+  try {
+    const body = await res.json();
+    return body?.detail ?? fallback;
+  } catch {
+    return fallback;
+  }
+}
+
+export async function deactivateUser(email: string): Promise<void> {
+  const res = await fetch("/api/manage/admin/deactivate-user", {
+    method: "PATCH",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ user_email: email }),
+  });
+  if (!res.ok) {
+    throw new Error(await parseErrorDetail(res, "Failed to deactivate user"));
+  }
+}
+
+export async function activateUser(email: string): Promise<void> {
+  const res = await fetch("/api/manage/admin/activate-user", {
+    method: "PATCH",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ user_email: email }),
+  });
+  if (!res.ok) {
+    throw new Error(await parseErrorDetail(res, "Failed to activate user"));
+  }
+}
+
+export async function deleteUser(email: string): Promise<void> {
+  const res = await fetch("/api/manage/admin/delete-user", {
+    method: "DELETE",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ user_email: email }),
+  });
+  if (!res.ok) {
+    throw new Error(await parseErrorDetail(res, "Failed to delete user"));
+  }
+}